| name | integrate-model |
| description | Integrate a new inference model into Xinity from a loose request like "install/integrate/add model X". Researches the model, assembles validated model data (the xinity-infoserver model entry), and verifies it actually runs via the daemon's run-model script, iterating on failures. Use whenever someone wants to make a specific model runnable on the cluster or to produce a model entry they can integrate themselves. |
Integrate a model
Follow the canonical procedure in packages/xinity-infoserver/docs/integrating-a-model.md:
research the model, assemble its model entry, verify it with run-model, and iterate on
failures using the failure/fix table. The field reference is in that doc and in
packages/xinity-infoserver/docs/model-fields.md.
When working through it as an agent:
- Write the model entry to a YAML file OUTSIDE this repo (e.g.
~/modeloutputs/<model>.yaml) and
point run-model --models <that file> at it for verification. NEVER edit or add to this repo's
packages/xinity-infoserver/models.yaml: model metadata is tracked externally and the integrator
decides where to apply what you produce. No model-data artifacts may land in this repository.
- Run
run-model with --json (--plan --json, then --start --json) so output is machine-readable;
branch on .gate.reason and, on error, the non-zero exit with .code, instead of scraping text.
- Confirm the entry by actually running it, not by reasoning alone, and that means a real request,
not just
/health.
- Research and validate declared capabilities: check whether the model supports tool/function
calling and vision, and if research says it plausibly does, add the tag (tools also needs
providerArgs.vllm: ["--tool-call-parser", "<name>"]) and test it against the running server (a
real tool_calls response; an image description). Skip a test only when research shows no realistic
chance, and note that you checked.
- Do not add
custom_code (vLLM --trust-remote-code) preemptively; only after a load failure shows it is needed.
- If the model is ambiguous (base vs instruct, size, quantization), ask the user before picking.
